Room: 5-314
Speaker Name:
Mengze Wang
Affiliation:
SandLab
Abstract:
Estimation of the probability of extreme events under climate change is both challenging and crucial for effective natural disaster risk management. Earth System Models (ESMs) offer detailed insights into climate extremes with high spatiotemporal resolution. However, the substantial computational cost of ESMs limits their use to projecting only a few future scenarios. Emulators, or reduced-complexity models, complement ESMs by enabling rapid predictions of local responses to diverse climate change scenarios. In this study, we introduce a data-driven framework to emulate the complete probability distribution of spatially-resolved climate extremes. We start with an emulator for the daily maximum temperature. The Empirical Orthogonal Functions (EOFs) are extracted to represent the dominant modes of the global climate system. The time-dependent coefficients of the leading EOFs are decomposed into long-term seasonal variations and daily fluctuations, with the former modeled as functions of the global mean temperature and the latter as stochastic processes. Our emulator is trained and validated using CMIP6 data. By generating multiple realizations, the emulator efficiently reproduces the temporal evolution of the weakly non-Gaussian probability distribution of local daily maximum temperatures. The accuracy of the emulated statistics in testing scenarios is consistent with the training datasets. We conclude by extending our framework to strongly non-Gaussian variables like precipitation using probabilistic diffusion models.